EDGE COLORINGS OF GRAPHS AVOIDING MONOCHROMATIC

MATCHINGS OF A GIVEN SIZE

CARLOS HOPPEN, YOSHIHARU KOHAYAKAWA, AND HANNO LEFMANN

Dedicated to the memory of Professor R. H. Schelp

Abstract. Let k and ` be positive integers. With a graph G, we associate the quantity
ck,` (G), the number of k-colorings of the edge set of G with no monochromatic matching of
size `. Consider the function ck,` : N −→ N given by ck,` (n) = max {ck,` (G) : |V (G)| = n},
the maximum of ck,` (G) over all graphs G on n vertices. In this paper, we determine ck,` (n)
and the corresponding extremal graphs for all large n and all fixed values of k and `.

1. Introduction
Let F be a fixed graph and k be a positive integer. In this paper, we study F -free k-colorings
of a graph G, that is, edge colorings of G with k colors such that there is no monochromatic
copy of F . (Edge colorings in this work need not be proper.) More precisely, given a graph
G, we consider the number ck,F (G) of F -free k-colorings of G, and we study the extremal
function ck,F : N −→ N, where ck,F (n) maximizes ck,F (G) over all graphs G on n vertices. In
other words, ck,F (n) = max {ck,F (G) : |V (G)| = n}. For instance, if there is a single color
available, we must have ck,F (n) = 1, with equality c1,F (n) = c1,F (G) for every graph G on n
vertices that does not contain a copy of F . A graph G on n vertices with ck,F (G) = ck,F (n) is
called (k, F )-extremal. The study of the function ck,F and of the set of (k, F )-extremal graphs
has been motivated by a conjecture of Erdős and Rothschild [3] concerning edge colorings of
graphs with a given number of colors and no monochromatic cliques of a given order.
The graphs F forbidden in this paper are matchings I` consisting of ` independent edges,
where ` ≥ 2. For convenience, we use the notation ck,I` (·) = ck,` (·) and we say that (k, I` )-
extremal graphs are simply (k, `)-extremal. To state the main result in this paper, we need a
preliminary definition.
Definition 1.1. The following concepts will be used throughout the paper.
(a) Given integers c ≥ 1 and n ≥ c + 2, let Gn,c = ([n], En,c ) be the graph on the vertex set
[n] = {1, . . . , n} such that {i, j} ∈ En,c if and only if min{i, j} ≤ c, i.e., the graph Gn,c
contains exactly c vertices each of which is joined by an edge to every other vertex.
(b) Given integers k, ` ≥ 2, let c(k, `) be the quantity defined by

`−1 if k ∈ {2, 3}
c(k, `) = (1)
d(` − 1)k/3e if k ≥ 4.
Theorem 1.2. Let k, ` ≥ 2 be fixed integers. There exists n0 = n0 (k, `) such that, for
n ≥ n0 , we have ck,` (n) = ck,` (Gn,c(k,`) ). Moreover, for n ≥ n0 , the graph Gn,c(k,`) is the
unique (k, `)-extremal graph up to isomorphism.
When ` = 2, this theorem is a very special case of a hypergraph result obtained by the
current authors in [6]. Erdős and Gallai [4] have shown that, for n sufficiently large, the
The first author acknowledges the support by FAPERGS (Proc. 10/0388-2) and CNPq (Proc. 484154/2010-
9). The second author was partially supported by CNPq (Proc. 308509/2007-2, 484154/2010-9). This work
was partially supported by the U. of São Paulo, through the MaCLinC project.
2 C. HOPPEN, Y. KOHAYAKAWA, AND H. LEFMANN

graph Gn,`−1 is the extremal graph of the forbidden graph I` , that is, Gn,`−1 is the graph on n
vertices with the largest number of edges that does not contain a matching I` as a subgraph.
Note that for k ∈ {2, 3} we have Gn,c(k,`) = Gn,`−1 . We should mention that, for general
forbidden graphs F , the number ex(n, F ) of edges in an n-vertex extremal graph for F is a
parameter that appears in natural upper and lower bounds on ck,F (n), namely
k ex(n,F ) ≤ ck,F (n) ≤ k k ex(n,F ) . (2)
To obtain the lower bound, note that an extremal graph G on n vertices for F has k ex(n,F )
distinct F -free k-colorings, as all k-colorings of the set of edges of G have this property.
Concerning the upper bound, for any k-coloring of the set of edges of a graph on n vertices
with k ex(n, F ) + 1 edges, at least one color class contains at least ex(n, F ) + 1 edges, and
hence contains a copy of F . Therefore the value of ck,F (n) is achieved by a graph with at
most k ex(n, F ) edges, from which we deduce that ck,F (n) ≤ k k ex(n,F ) .
Theorem 1.2 implies that, for large n, the lower bound in (2) is tight for k ∈ {2, 3},
whereas it is not tight for k ≥ 4. This phenomenon replicates, for the forbidden graphs I` ,
what has been observed in other classes of forbidden graphs F , such as complete graphs and
odd cycles (see Yuster [13] and Alon, Balogh, Keevash and Sudakov [1], which rely on the
Szemerédi Regularity Lemma [12]). In the case of complete graphs, this verifies the Erdős
and Rothschild Conjecture for k ∈ {2, 3}, but disproves it for k ≥ 4. More generally, Lemma
2.1 in [1] implies that, given a graph F and k ∈ {2, 3}, we have
2
ck,F (n) ≤ k ex(n,F )+o(n ) .
In other words, the extremal graph of F is never “far” from being (k, F )-extremal when k ∈
{2, 3} and F is not bipartite, as the Erdős-Stone Theorem [5] ensures that ex(n, F ) = Ω(n2 ) in
this case. The work in [1] also implies that ck,F (n) > k ex(n,F ) if F is not bipartite and k ≥ 4.
To the best of our knowledge, this is the first time that the same behavior is observed for a
class of bipartite graphs.
Naturally, one might ask whether the following holds for every fixed graph F : given k ∈
{2, 3} and ε > 0, there exists n0 such that
ck,F (n) ≤ k (1+ε) ex(n,F )
for every n ≥ n0 . However, the current authors [7] have answered this question in the negative.
As a matter of fact, the extremal graph is far from optimal for 3-colorings with forbidden
paths with two edges (i.e., for proper 3-edge colorings). The same occurs for 2-colorings with
forbidden stars on t ≥ 3 leaves.
In the last few years, there has been substantial progress in the study of ck,F (n) and of
closely-related graph and hypergraph functions. For instance, in the case of graphs, Pikhurko
and Yilma [11] have determined, for n sufficiently large, the families of graphs G on n vertices
such that c4,K3 (G) = c4,K3 (n) and such that c4,K4 (G) = c4,K4 (n). Balogh [2] has considered
another question, which extends the problem of finding ck,F (n). For a fixed graph F and a
fixed edge coloring of F , he wishes to determine the n-vertex graph with the largest number of
edge colorings avoiding copies of F with the prescribed coloring. He has shown that, for any
2-coloring of the edges of a complete graph Kr , where r ≥ 3, maximality with respect to this
problem is only achieved by the extremal graph for Kr (namely the Turán graph Tr−1 (n)), as
long as n is sufficiently large.
Concerning hypergraphs, recent results by Lefmann, Person, Rödl and Schacht [9] estab-
lished the following: if k ∈ {2, 3}, F is the Fano plane and n is sufficiently large, then the
number of k-colorings of the hyperedges of any n-vertex 3-uniform hypergraph H on n vertices
is at most k ex(n,F ) , where, as usual, the quantity ex(n, F ) denotes the maximum number of
hyperedges in a 3-uniform n-vertex hypergraph containing no Fano plane. They have also
 EDGE COLORINGS AVOIDING MONOCHROMATIC MATCHINGS OF A GIVEN SIZE 3

shown that equality is attained by the unique extremal hypergraph for ex(n, F ) and that,
for fixed k ≥ 4, the inequality ck,F (n) > k ex(n,F ) holds as n tends to infinity. Very recently,
a similar phenomenon has been proved to hold in several other instances, such as 3- and
4-uniform generalized triangles, expanded complete 2-graphs and Fan(k)-hypergraphs. For
more information, see Lefmann and Person [8] and Lefmann, Person and Schacht [10].
This paper is organized as follows. Section 2 contains the proof of Theorem 1.2 when k = 2
colors are used. The structure of the proof of this result for k ≥ 3 is delineated in Section 3,
leaving the proofs of technical tools to the subsequent sections.

2. Proving Theorem 1.2 for k = 2

Our objective in this section is to determine the value of c2,` (n) for every fixed ` and every
sufficiently large n. If n ≤ 2` − 1, then no graph on n vertices can contain a matching I` of
n
size `, hence c2,` (n) = 2( 2 ) and the only graph to attain this extremal value is the complete
graph Kn . The following result, due to Erdős and Gallai [4], fully describes the extremal
graphs for I` .
Theorem 2.1. Given integers n ≥ 2` ≥ 4, we have
   
2` − 1 `−1
ex(n, I` ) = max , (` − 1)(n − ` + 1) + .
2 2
Equality can occur only if G is the union of a complete graph on 2` − 1 vertices with n − 2` + 1
isolated vertices, or if G has a vertex cover C of size ` − 1 and contains all possible edges with
at least one vertex in C.
In our arguments we shall make use of the following easy result.
Lemma 2.2. Let G = (V, E) be a bipartite graph with bipartition V = A ∪ B, where |A| = `.
If every vertex v ∈ A has degree at least `, then G contains a matching I` .
Next, we shall prove the following, which is essentially the statement of Theorem 1.2 in the
case when k = 2.
Lemma 2.3. For ` ≥ 2, there exists n0 = n0 (`) such that, for n ≥ n0 , we have
`−1
c2,` (n) = 2ex(n,I` ) = 2(`−1)(n−`+1)+(2 ).

In addition to this, if n ≥ n0 , an n-vertex graph G satisfies c2,` (G) = c2,` (n) if and only if G
is isomorphic to the graph Gn,`−1 (see Definition 1.1). Moreover, if a graph G on n vertices
contains a matching I` , then c2,` (G)  c2,` (n); in fact, c2,` (G) is exponentially smaller than
c2,` (n).
Proof. Let ` ≥ 2 be fixed. Throughout the proof, we shall assume that n is sufficiently large
for the bounds to hold, and no special attempt is made to optimize n0 . In particular, we
assume that n satisfies
5
`−1<n
2
4`−4
2( 2 )+4`−4 · `2 · n4(`−1)`  2n−`+1 .
The first condition ensures that Gn,`−1 is, up to isomorphism, the unique extremal graph for
I` , while the second naturally arises in a counting argument.
It is clear that, for every value of n and every I` -free n-vertex graph G = (V, E), we have
c2,` (G) = 2|E| ≤ 2ex(n,I` ) , as no 2-coloring of an I` -free n-vertex graph could possibly contain
a monochromatic I` . Moreover, equality can be attained by extremal I` -free n-vertex graphs,
4 C. HOPPEN, Y. KOHAYAKAWA, AND H. LEFMANN

`−1
so that c2,` (n) ≥ 2ex(n,I` ) = 2(`−1)(n−`+1)+( 2 ) for n large enough, where the equality sign
holds by Theorem 2.1 and our assumption on n.
Now, let G be an n-vertex graph that admits I` -free 2-colorings of its set of edges, but
contains I` as a subgraph. Let Im be a maximum matching in G given by the edges e1 =
{a1 , b1 }, . . . , em = {am , bm }, m ≥ `. We shall see that the number of I` -free 2-colorings of G
is exponentially smaller than 2ex(n,I` ) . Clearly, we must have m ≤ 2(` − 1), as otherwise the
matching itself would not be 2-colorable. Moreover, the maximality of Im implies that the set
W of vertices uncovered by the matching forms an independent set in G, as any edge between
vertices in W could be used to increase Im to a matching Im+1 .
2m
Let A = {a1 , . . . , am } and B = {b1 , . . . , bm }. Let ∆ be one of the at most 2( 2 ) I` -free 2-
colorings of the set of edges with both endpoints in A∪B. Assume that the colors used are red
and green and w.l.o.g. suppose that, with respect to ∆, the edges e1 , . . . , er are red, and the
edges er+1 , . . . , em are green. By assumption, we have 1 ≤ r ≤ `−1 and 1 ≤ g = m−r ≤ `−1.
Our objective is to bound the number of possible extensions of ∆ to an I` -free 2-coloring
of G. To achieve this, we look at the structure of the coloring ∆ more closely. Given a vertex
v ∈ V , we classify its neighbors as red or green according to the color of the edge joining v to
each neighbor.
Claim 2.4. Let ∆0 be an I` -free extension of the partial coloring ∆ to G.
(a) The number of vertices in the set {ar+1 , br+1 , . . . , am , bm } with ` − r or more red
neighbors in W is bounded above by ` − r − 1.
(b) The number of vertices in the set {a1 , b1 , . . . , ar , br } with ` − g or more green neighbors
in W is bounded above by ` − g − 1.
Proof. For a contradiction, assume instead that there is a subset S ⊂ {ar+1 , br+1 , . . . , am , bm },
|S| ≥ ` − r, such that every vertex in S has at least ` − r red neighbors in W . By Lemma 2.2
there must be a matching of size ` − r in ∆0 formed uniquely by red edges with one endpoint
in S and the other in W . Together with the edges e1 , . . . , er , this yields a red matching I` ,
which is not possible.
Analogously, the number of vertices in the set {a1 , b1 , . . . , ar , br } with at least ` − (m −
r) = ` − g green neighbors in W with respect to some extension of ∆ is bounded above by
` − (m − r) − 1 = ` − g − 1. This asserts the validity of the above claim. 
So, to extend ∆ to an I` -free 2-coloring of G, one may first choose j1 (0 ≤ j1 ≤ ` − g − 1)
vertices in {a1 , b1 , . . . , ar , br } to have at least ` − g green neighbors in W , and j2 (0 ≤ j2 ≤
` − r − 1) vertices in {ar+1 , br+1 , . . . , am , bm } to have at least ` − r red neighbors in W . Once
these vertices are chosen, there are at most 2(j1 +j2 )(n−2m) ways of coloring the set of edges
joining them to W . The set of edges between each of the 2m − j1 − j2 remaining vertices in
{a1 , b1 , . . . , am , bm } and W may be colored in at most
max{`−r−1,`−g−1}  
X n − 2m
i
i=0
ways. Hence, the total number of extensions of ∆ to an I` -free 2-coloring of G is bounded
above by
 2m−j1 −j2
 
`−g−1 max{`−r−1,`−g−1} 
X `−r−1
X 2r2m − 2r X n − 2m 
2(j1 +j2 )(n−2m) 
j1 j2 i
j1 =0 j2 =0 i=0
`−g−1
X `−r−1
X   
2r 2m − 2r `(2m−j1 −j2 ) (j1 +j2 )(n−2m)
≤ n 2
j1 j2
j1 =0 j2 =0
 EDGE COLORINGS AVOIDING MONOCHROMATIC MATCHINGS OF A GIVEN SIZE 5

< `2 · 22m · n2m` · 2(2`−r−g−2)(n−2m)

= `2 · 22m · n2m` · 2(2`−m−2)(n−2m) .
2m
Summing over all of the at most 2( 2 ) possible choices of ∆ and observing that ` ≤ m ≤
2(` − 1), we conclude that, for n sufficiently large, c2,` (G) is bounded above by
2m
2( 2 ) · `2 · 22m · n2m` · 2(2`−m−2)(n−2m)
4`−4
≤ 2( 2 ) · 24`−4 · `2 · n2m` · 2(2`−m−2)(n−2m)
4`−4
< 2( 2 )+4`−4 · `2 · n4(`−1)` · 2(`−2)(n−`+1) (3)
 2(`−1)(n−`+1) < 2ex(n,I` ) .
Note that (3) is exponentially smaller than 2ex(n,I` ) , which finishes the proof. 

3. Proof of Theorem 1.2 for k ≥ 3

In this section, we describe the proof of Theorem 1.2 in the case k ≥ 3. For convenience,
we use the term (k, `)-coloring of a graph G to mean an I` -free k-coloring of the set of edges
of G. A graph G is (k, `)-colorable if it admits a (k, `)-coloring.
We frequently use the following lemma, whose simple proof can be found in [6].
Lemma 3.1. Let m ≥ 2 be an integer. All optimal solutions s = (s1 , . . . , sc ) to the maxi-
mization problem
max ci=1 sc
Q
c, s1 , . . . , sc positive integers, (4)
s1 + · · · + sc ≤ m,
have the following form.
(a) If m ≡ 0 (mod 3), then c = m/3 and all the components of s are equal to 3.
(b) If m ≡ 1 ( mod 3), then either c = dm/3e, with exactly two components equal to 2 and
all remaining components equal to 3, or c = bm/3c, with exactly one component equal
to 4 and all remaining components equal to 3.
(c) If m ≡ 2 (mod 3), then c = dm/3e with exactly one component equal to 2 and all
remaining components equal to 3.
As a consequence, the optimal value of (4) is 3m/3 if m ≡ 0 (mod 3), and 4 · 3bm/3c−1 if
m ≡ 1 (mod 3), and 2 · 3bm/3c if m ≡ 2 (mod 3).
Given the number k of colors and the forbidden matching I` , the following function appears
often in the remainder of this paper.
Definition 3.2. Given positive integers k and `, let D(k, `) be defined by
 (`−1)k/3
 3 if (` − 1)k ≡ 0 (mod 3)
D(k, `) = 4 · 3b(`−1)k/3c−1 if (` − 1)k ≡ 1 (mod 3)
2 · 3b(`−1)k/3c if (` − 1)k ≡ 2 (mod 3).


Let 0 < γ = γ(k, `) < 1/3Pbe such that, whenever the nonzero components of an integral
vector s = (s1 , . . . , sc ) with ci=1 si ≤ (` − 1)k and si ≤ k for every i do not form an optimal
solution to (4) with m = (` − 1)k, we have
Yc
si < D(k, `)1−3γ . (5)
i=1,si 6=0
For the proof of Theorem 1.2, we shall now establish the following result.
Lemma 3.3. Let k ≥ 3 and ` ≥ 2 be integers. Then there exists an integer n0 such that, if
n ≥ n0 , we have ck,` (n) = ck,` (Gn,c(k,`) ). Moreover, the following assertions hold:
6 C. HOPPEN, Y. KOHAYAKAWA, AND H. LEFMANN

(i) the graph Gn,c(k,`) is, up to isomorphism, the unique (k, `)-extremal graph;
(ii) in particular, for k = 3, we have
`−1
c3,` (n) = 3ex(n,I` ) = 3(`−1)(n−`−1)+( 2 );
(iii) for (` − 1)k 6≡ 1 (mod 3), if a graph G on n vertices has a minimum vertex cover of
size c 6= c(k, `), then ck,` (G) is exponentially smaller than ck,` (n).
Proof. Let k ≥ 3 and ` ≥ 2 be fixed. We first prove the result for (` − 1)k 6≡ 1 ( mod 3). With
foresight, choose n0 sufficiently large so that, for every n ≥ n0 ,
k k(`−1) D(k, `)n(1−γ) < D(k, `)n−c(k,`) ,
2(`−1)k
k ( 2 ) < D(k, `)γn , and (6)
2 2
(2`k)(`−1)k · e2` k (ln n+ln k) < D(k, `)γn .
The constant γ = γ(k, `) in these inequalities satisfies the condition stated in (5).
Let G = (V, E) be a (k, `)-colorable graph on n ≥ n0 vertices. The proof is structured in
terms of a minimum vertex cover C = {v1 , . . . , vc } of G. The first step is to obtain an upper
bound on the size c = |C| that does not depend on n.
Lemma 3.4. If G = (V, E) is a (k, `)-colorable graph, then G has a vertex cover of size at
most 2k(` − 1).
Proof. Let F be a maximum set of pairwise independent edges in G and let ∆ be a (k, `)-
coloring of G. Since there are at most k color classes in ∆ and the intersection of F with each
color class has size at most ` − 1, we conclude that |F | ≤ (` − 1)k. Due to the maximality of
F , we deduce that the endpoints of the edges in F form a vertex cover of G, and our result
follows. 
The second step of the proof is to show that, if G has the largest number of (k, `)-colorings
over all n-vertex graphs, then G must have a minimum vertex cover of size c(k, `).
Lemma 3.5. For (` − 1)k 6≡ 1 (mod 3), if n ≥ n0 and an n-vertex graph G = (V, E) is
(k, `)-extremal, then the minimum size of a vertex cover in G is equal to c(k, `).
To conclude the proof, we show that Gn,c(k,`) is the unique (k, `)-extremal graph over all
graphs on n vertices with minimum vertex cover of size c(k, `).
Lemma 3.6. For (` − 1)k 6≡ 1 (mod 3), up to isomorphism, the unique maximum of ck,` (G)
over all graphs G with vertex set of size n ≥ n0 and minimum vertex cover of size c(k, `) is
achieved by the graph Gn,c(k,`) .
The proofs of Lemmas 3.5 and 3.6 are given in Sections 4 and 5, respectively.
We now comment on the case (` − 1)k ≡ 1 (mod 3). The main difference is that, in this
case, there are two families of optimal solutions to (4) with m = (` − 1)k. Because of this, the
arguments used in the proof of Lemma 3.5 only show that, for large n, the minimum size of
the vertex cover in a (k, `)-extremal graph G, i.e., a graph for which ck,` (G) = ck,` (n), lies in
the set {c(k, `) − 1, c(k, `)}. We may then adapt the arguments in the proof of Lemma 3.6 to
demonstrate that the (k, `)-extremal graph is, up to isomorphism, either Gn,c(k,`) or Gn,c(k,`)−1 .
To conclude the proof, we then show that the latter has fewer (k, `)-colorings than Gn,c(k,`) .
Details regarding these steps may be found in Section 6. 

4. Proof of Lemma 3.5

We start this section with a lower bound on the number of (k, `)-colorings of the graph
Gn,c(k,`) . Let C = {v1 , . . . , vc(k,`) } be a minimum vertex cover of this graph, consider an opti-
mal solution s = (s1 , . . . , sc(k,`) ) to the optimization problem (4), and fix a way of distributing
 EDGE COLORINGS AVOIDING MONOCHROMATIC MATCHINGS OF A GIVEN SIZE 7

k(` − 1) balls, with (` − 1) balls having each of the k colors, into c(k, `) distinct bins in such
a way that the i-th bin receives si balls, all of which have different colors. Assuming that
the colors are indexed by 1, . . . , k, we consider the set S of all edge colorings of Gn,c(k,`) with
the following properties: (i) if e = {vi , vj }, 1 ≤ i < j ≤ c(k, `), then e is colored with the
color of smallest index among all colors in bin i or j; (ii) if e = {vi , w}, where w ∈ / C, then e
may be colored with any of the colors in bin i. It is easy to see that every coloring in S is a
(k, `)-coloring of Gn,c(k,`) . Hence we have
c(k,`)
n−c(k,`)
Y
ck,` (Gn,c(k,`) ) ≥ |S| = si = D(k, `)n−c(k,`) . (7)
i=1
We shall see that graphs whose minimum vertex covers do not have size c(k, `) have fewer
(k, `)-colorings than D(k, `)n−c(k,`) when n ≥ n0 . To this end, let G = (V, E) be a (k, `)-
colorable graph on n vertices with minimum vertex cover C = {v1 , . . . , vc }. Let F be the set
of all edges of G with both endpoints in C and consider the graph G0 obtained from G by the
removal of F . Note that G0 is a bipartite graph with bipartition (C, V \ C).
Consider a (k, `)-coloring ∆ of G. For a vertex vi ∈ C and a color σ ∈ {1, . . . , k}, we say
that σ is substantial for vi if at least ` edges incident with vi in G0 have color σ. In other
words, a color is substantial for vertex vi if it appears “many” times among the edges that
are incident with vi , but not with any other cover element Moreover, we say that vertex vi is
s-influential if there are precisely s colors that are substantial for vi . Finally, a color σ is said
to be influential if it is substantial for exactly ` − 1 vertices in C. Observe that influential
colors are “maximally substantial”, since a color cannot be substantial for more than ` − 1
vertices due to Lemma 2.2.
Given j ∈ {0, . . . , (` − 1)k}, let Ij be the set of all non-negative integral solutions to the
equation s1 + · · · + sc = j such that si ≤ k for every i. For any such vector s = (s1 , . . . , sc ),
let ∆s (G0 ) be the set of all (k, `)-colorings of G0 for which vertex vi is si -influential for every
i ∈ {1, . . . , c}. As the sets ∆s (G0 ) partition the set of all (k, `)-colorings of G0 , we have
(`−1)k
X X
0
ck,` (G ) = ∆s (G0 ) . (8)
j=0 s∈Ij

Clearly, every (k, `)-coloring of G is the combination of a (k, `)-coloring of G0 with a coloring
of the edges of F with at most k colors. We know that there are at most
c
k |F | ≤ k (2) (9)
colorings of the latter type, thus equation (8) becomes
(`−1)k
c
ck,` (G) ≤ k (2)
X X
∆s (G0 ) . (10)
j=0 s∈Ij

We now concentrate on (k, `)-colorings of G0 .

Lemma 4.1. If a color σ is influential with respect to a (k, `)-coloring ∆, then every edge e
with color σ is incident with a vertex for which it is substantial.
Proof. Suppose that σ is influential with respect to ∆ and let vi1 , . . . , vi`−1 be vertices in
the vertex cover C for which σ is substantial. Let e = {x, vj } be an edge in G0 , where
vj ∈ C \ {vi1 , . . . , vi`−1 } and x ∈
/ C, and assume that e has color σ. Consider the bipartite
subgraph B of G0 induced by U = {vi1 , . . . , vi`−1 } and by the set W of all vertices y 6= x that
are adjacent to a vertex in U through an edge with color σ. The degree of every vertex in
U is at least ` − 1 in B, hence B contains a matching M of size ` − 1 by Lemma 2.2. Thus
8 C. HOPPEN, Y. KOHAYAKAWA, AND H. LEFMANN

M ∪ {e} is also a matching, and every edge in this matching has color σ, contradicting the
assumption that ∆ is a (k, `)-coloring of G. 

Lemma 4.1 will be used to bound from above the number of colorings in ∆s (G0 ). Fix a
vector s = (s1 , . . . , sc ) with s1 + · · · + sc = j, where si ≤ k for every i. The number of ways
to choose
the colors that are substantial, with their respective multiplicity, is bounded above
by (`−1)k
j , which is the number of ways of choosing j different colors from a set of (` − 1)k
different colors. Once the colors are chosen, they may be distributed in at most s1 !s2j!!···sc ! ways
among the cover vertices, where the upper bound again pretends that all the colors chosen are
distinct. Now, given that the colors have been distributed, the edges in G0 that are incident
with vertex vi may be colored in at most
k−s k−s
! !
X Yi n − c n−c−Pk−si at X Yi n − c
si t=1
≤ sin−c
at at
(a1 ,...,ak−si ) t=1 (a1 ,...,ak−si ) t=1

ways, if si ≥ 1, where the sum is such that each at ranges from 0 to `−1. This is because there
are at most n − c edges in G0 that are incident with vi , from which at , 0 ≤ at ≤ ` − 1, may be
chosen for every color that is not substantial for vertex vi . All remaining edges incident with
vertex vi may be colored with any of the si colors that are substantial for vi . Note that
k−s k−s
! !
X Yi n − c X Yi
n−c
si ≤ (n − c) at
sn−c
i
at
(a1 ,...,ak−si ) t=1 (a1 ,...,ak−si ) t=1
   k−si
X Pk−si `−1
X
at  n−c
=  (n − c) t=1 si =  (n − c)p  sn−c
i
(a1 ,...,ak−si ) p=0
k−si
(n − c)` − 1

= sn−c
i ≤ n`k · sin−c . (11)
n−c−1
If si = 0, then there are at most (` − 1)k edges incident with vertex vi in G0 , which may
be colored with at most k colors in at most

k (`−1)k (12)

ways.
Combining inequalities (11) and (12), and observing that c is an upper bound on the number
of vanishing components in a vector s = (s1 , . . . , sc ), we obtain
  c
X
0 (` − 1)k X j! Y
∆s (G ) ≤ c`k
·n ·k c(`−1)k
· sin−c
j s1 !s2 ! · · · sc !
s∈Ij s∈Ij i=1,si 6=0
  c
(` − 1)k X j! Y
≤ · ec`k(ln n+ln k) · sn−c
i . (13)
j s1 !s2 ! · · · sc !
s∈Ij i=1,si 6=0

Observe that, for our fixed value of k, the product ci=1,si 6=0 sin−c is maximized when the
Q

nonzero components of s are the components of a vector in the set S(k, `) of optimal solutions
to (4) with m = k(` − 1), which is described in Lemma 3.1. Recall that the number D(k, `)
given in the statement of Definition 3.2 is precisely the optimal value of (4), and, whenever
the nonzero components of the integral vector s = (s1 , . . . , sc ) are not an optimal solution to
 EDGE COLORINGS AVOIDING MONOCHROMATIC MATCHINGS OF A GIVEN SIZE 9

(4), we have
c
Y
si < D(k, `)1−3γ . (14)
i=1,si 6=0

Lemma 4.2. Let k and ` be positive integers and fix n0 as in (6). For n ≥ n0 , let G be an
n-vertex graph with minimum vertex cover of cardinality c ≤ 2(` − 1)k. Then
k(`−1)
c
k ( 2)
X X
∆s (G0 ) ≤ D(k, `)n(1−γ) .
j=0 s∈Ij \S(k,`)

In particular, if ∆s (G0 ) = ∅ for every s = (s1 , . . . , sc ) whose nonzero components are the
components of a vector in the set S(k, `) of optimal solutions to (4), then

ck,` (G0 ) ≤ D(k, `)n(1−γ) .

Proof. Let G be an n-vertex graph with a minimum vertex cover C of cardinality c ≤ 2k(` −
1) < 2k`. Each edge with both endpoints in the vertex cover C can be colored with at most
k colors. The inequalities (13) and (14) imply with (6) that
(`−1)k
c
k (2) ·
X X
∆s (G0 )
j=0 s∈Ij \S(k,`)
(`−1)k  
2(`−1)k (` − 1)k j!
≤ k( 2 ) ·
X X
· ec`k(ln n+ln k) · D(k, `)(n−c)(1−3γ)
j s1 !s2 ! · · · sc !
j=0 s∈Ij
(`−1)k   X
2(`−1)k (` − 1)k j!
= k ( 2 ) · ec`k(ln n+ln k) · D(k, `)(n−c)(1−3γ) ·
X
·
j s1 !s2 ! · · · sc !
j=0 s∈Ij
(`−1)k  
X (` − 1)k j
≤ D(k, `)γn · ec`k(ln n+ln k) · D(k, `)(n−c)(1−3γ) · c
j
j=0
2`2 k2 (ln n+ln k)
≤ e · D(k, `)n(1−2γ) · (1 + c)(`−1)k
2 k 2 (ln n+ln k)
≤ (2`k)(`−1)k · e2` · D(k, `)n(1−2γ)
≤ D(k, `)γn · D(k, `)n(1−2γ) = D(k, `)n(1−γ) , (15)
as required.
When ∆s (G0 ) = ∅ for every s = (s1 , . . . , sc ) whose nonzero components are the components
of a vector in S(k, `), the inequality ck,` (G) ≤ D(k, `)n(1−γ) is an immediate consequence of
the above. 

In light of Lemma 4.2, to conclude the proof of Lemma 3.5, it suffices to show the following.
Claim 4.3. If c 6= c(k, `), then ∆s (G0 ) = ∅ for every s = (s1 , . . . , sc ) whose nonzero compo-
nents are the components of a vector in S(k, `).
Indeed, if this claim is true, then by Lemma 4.2 we have
ck,` (G) ≤ D(k, `)n(1−γ)  D(k, `)n−c(k,`) ≤ ck,` (Gn,c(k,`) ),
which implies the statement of Lemma 3.5.
10 C. HOPPEN, Y. KOHAYAKAWA, AND H. LEFMANN

Proof. Claim 4.3 is certainly true for c < c(k, `), as there are no vectors of length c whose
nonzero components give a solution in S(k, `). To prove the claim for c > c(k, `), we note
that Lemma 4.1 implies that, if ∆ is a (k, `)-coloring of G for which every color is influential,
then, for every vertex vi in the minimum vertex cover C, there is a color that is substantial
for vi . Indeed, if we suppose for a contradiction that this is not true, then Lemma 4.1 tells
us that there is no color available for edges incident with vertex vi in G0 ; however, such edges
must exist by the minimality of C.
With this, we conclude that ∆s (G0 ) = ∅ for every s = (s1 , . . . , sc ) with s1 +· · ·+sc = (`−1)k
but sj = 0 for some j, which in turn proves our result. 

5. Proof of Lemma 3.6

By Lemma 3.5, for n ≥ n0 , an n-vertex graph G = (V, E) is (k, `)-extremal only if the size
of a minimum vertex cover of G is c(k, `). Thus let G = (V, E) be an n-vertex graph with
minimum vertex cover C = {v1 , . . . , vc(k,`) } for which there are vertices vi ∈ C and x 6= vi
such that {vi , x} ∈
/ E. We shall show that the graph G is not (k, `)-extremal. This is done in
two parts. First, we prove that every (k, `)-coloring of G can be extended to a (k, `)-coloring
of G ∪ {vi , x}. To conclude the proof, we show that at least one of the (k, `)-colorings of G
can be extended in more than one way, and hence ck,` (G) < ck,` (G ∪ {vi , x}).
We first consider the case when vertex vi covers at most (` − 1)k edges not covered by any
other element of C. Consider the graph G0 obtained from G by the removal of these edges.
On the one hand, since G0 has a vertex cover of size smaller than c(k, `), the number of (k, `)-
colorings of G0 is bounded above by D(k, `)n(1−γ) . On the other hand, every (k, `)-coloring of
G consists of the union of a (k, `)-coloring of G0 and a k-coloring of the edges removed from
G, so that there are at most
k k(`−1) D(k, `)n(1−γ) (16)
such colorings, which is smaller than ck,` (Gn,c(k,`) ) ≥ D(k, `)n−c(k,`) by our choice of n0 in (6).
Now, suppose that vertex vi covers more than (` − 1)k edges not covered by any other
element of C. Then, by the pigeonhole principle, for every (k, `)-coloring ∆ of G, there is a
color σ that is substantial for vertex vi . If ∆ induces a matching M of size ` − 1 with color
σ, then, as σ is substantial for vi , the vertex vi must be an endpoint of some edge in M . In
other words, the coloring ∆ can be extended to a (k, `)-coloring of G ∪ {vi , x} by assigning
the color σ to {vi , x}.
To conclude the proof, we observe that there exists a coloring ∆ of G using dc(k, `)/(` − 1)e
colors, namely the one that assigns the first color to all edges incident with the first ` − 1
cover elements, then assigns the second color to all uncolored edges incident with the next
` − 1 cover elements, and so on. However, note that
   
c(k, `) ((` − 1)k + 2)/3 ((` − 1)k + 2)/3 ` − 2
≤ ≤ + <k
`−1 `−1 `−1 `−1
if k ≥ 2. Hence the coloring ∆ can be extended to a coloring of G ∪ {vi , x} in at least two
ways, either by coloring {vi , x} with a color that is substantial for vertex vi or by coloring
{vi , x} with a color that is not used by ∆.

6. The case (` − 1)k ≡ 1 (mod 3)

To finish the proof of Theorem 1.2, we still need to address the case (` − 1)k ≡ 1 (mod 3).
The peculiarity of this case lies in the fact that there are two types of optimal solutions to the
optimization problem (4). As we have discussed in Section 3, this implies that the arguments
used in the proof of Lemma 3.5 only allow one to show that, for large n, the minimum size of
the vertex cover in a (k, `)-extremal graph G, i.e., a graph for which ck,` (G) = ck,` (n), lies in
 EDGE COLORINGS AVOIDING MONOCHROMATIC MATCHINGS OF A GIVEN SIZE 11

the set {c(k, `) − 1, c(k, `)}. With this, it is not hard to adapt the arguments in the proof of
Lemma 3.6 to show that the (k, `)-extremal graph lies in the set {Gn,c(k,`)−1 , Gn,c(k,`) }. We
then show that maximality is attained by Gn,c(k,`) .
Lemma 6.1. For (` − 1)k ≡ 1 ( mod 3), up to isomorphism, the maximum of ck,` (G) over all
graphs G with vertex set of size n ≥ n0 and minimum vertex cover of size either c(k, `) − 1
or c(k, `) is achieved by Gn,c(k,`)−1 or Gn,c(k,`) .
Proof. Let G be a graph with minimum vertex cover of size c(k, `) − 1 and assume that
G 6= Gn,c(k,`)−1 . The fact that ck,` (G) < ck,` (Gn,c(k,`)−1 ) can be derived directly from the
proof of Lemma 3.6, unless k = 4 and ` = 2, in which case c(k, `) − 1 = 1 and the conclusion
is straightforward.
We now show that ck,` (G) < ck,` (Gn,c(k,`) ) for every graph G 6= Gn,c(k,`) with minimum
vertex cover of size c(k, `). Let C = {v1 , . . . , vc(k,`) } be a minimum vertex cover of G and
assume that there are vertices vi ∈ C and x 6= vi such that {vi , x} ∈ / E. We shall show that
the graph G is not (k, `)-extremal. Let m be the number of edges whose only endpoint in C is
vi (the minimality of C implies that m ≥ 1). As in the proof of Lemma 3.6, we consider two
cases, according to whether m ≤ (` − 1)k or m > (` − 1)k. In the latter case, we may again
extend colorings of G to colorings of G ∪ {vi , x}, and the result follows with the arguments
used in the case (` − 1)k 6≡ 1 (mod 3).
For m ≤ (` − 1)k, let G0 be obtained from G by the deletion of all edges whose single
endpoint in C is vi . In this case, we need to be more careful than in the proof of Lemma 3.6,
as the upper bound (16) does not hold directly for the number of (k, `)-colorings in G0 , since,
when (` − 1)k ≡ 1 ( mod 3), there is an optimal solution to (4) of length c(k, `) − 1. Let C1 be
the family of (k, `)-colorings of G for which there is a color that is substantial for vi , and let
C2 contain the remaining (k, `)-colorings of G, so that ck,` (G) = |C1 | + |C2 |. We shall bound
the number of (k, `)-colorings in each of these two families separately.
Let ∆ be a coloring in C1 , and consider the coloring ∆0 of G0 obtained by ignoring the
colors of the deleted edges. Let C10 be the set of all (k, `)-colorings of G0 obtained in this way.
If σ is substantial for vi with respect to ∆, then it can be substantial for at most ` − 2 other
vertices in C, and hence σ cannot be influential with respect to ∆0 . By Lemma 4.2 applied
to G0 , we have
|C10 | ≤ D(k, `)n(1−γ) .
Moreover, each coloring of C10 corresponds to at most k k(`−1) − 1 colorings of C1 , since there
is at least one coloring of the deleted edges for which no color is substantial (namely one in
which there are at most ` − 1 edges with each of the k colors). In particular, we have
   
|C1 | ≤ k k(`−1) − 1 |C10 | ≤ k k(`−1) − 1 D(k, `)n(1−γ) .
Now, let ∆ be a coloring in C2 and let σ be the color of an edge whose single endpoint in
the vertex cover is vi . Since σ is not substantial for vi , Lemma 4.1 implies that σ cannot be
influential with respect to ∆, so that, by applying Lemma 4.2 to G, we obtain
|C2 | ≤ D(k, `)n(1−γ) .
As a consequence, we have
ck,` (G) = |C1 | + |C2 | ≤ k k(`−1) D(k, `)n(1−γ) < ck,` (Gn,c(k,`) ),
as required.

To conclude the analysis of the case (` − 1)k ≡ 1 (mod 3), we show that maximality is
attained by Gn,c(k,`) .
12 C. HOPPEN, Y. KOHAYAKAWA, AND H. LEFMANN

Lemma 6.2. For (` − 1)k ≡ 1 ( mod 3) and large n, we have ck,` (Gn,c(k,`) ) > ck,` (Gn,c(k,`)−1 ).
Proof. For simplicity, we use the notation c = c(k, `). Let C ∗ = {v1 , . . . , vc−1 } be the minimum
vertex cover of G∗ = Gn,c−1 and let vc be an additional vertex in this graph. We may view
G = Gn,c as the graph obtained by the addition of all edges {vc , w} to G∗ for every vertex
w∈/ C ∗.
Let S1 = S1 (k, `) and S2 = S2 (k, `) denote the sets of optimal solutions to the optimization
problem (4) with one component equal to 4 and with two components equal to 2, respectively.
From equation (8), combined with Lemma 4.2, we deduce that
X X
|∆s (G)| ≤ ck,` (G) ≤ |∆s (G)| + D(k, `)n(1−γ)
s∈S2 s∈S2
X X
|∆s∗ (G∗ )| ≤ ck,` (G∗ ) ≤ |∆s∗ (G∗ )| + D(k, `)n(1−γ) .
s∗ ∈S1 s∗ ∈S1

Because of equation (7), we know that, for large n, the two inequalities are dominated by the
sums. Hence, in order to establish our result, it suffices to show that
X X
|∆s (G)| ≥ 2 |∆s∗ (G∗ )| (17)
s∈S2 s∗ ∈S1
for n sufficiently large.
Note that, for every vector s∗ in S1 and every coloring in ∆s∗ (G∗ ), there exists exactly one
vertex vi ∈ C ∗ with the property that precisely four colors are substantial for vi . Moreover,
these are the only colors that appear in the edges for which vi is the single endpoint in C ∗ .
This will be used to define an injective map φ of the colorings of ∆s∗ (G∗ ) into ∆s (G), where
s is the vector in S2 whose components equal to 2 are precisely si and sc .
To define the mapping φ, suppose that the k colors are ordered from 1 to k, fix a coloring
∆∗ ∈ ∆s∗ (G∗ ) with s∗i = 4, and let σ1 < σ2 < σ3 < σ4 be the substantial colors in vi .
The image of ∆∗ under φ is a coloring ∆ ∈ ∆s (G) with σ1 , σ2 substantial for vi and σ3 , σ4
substantial for vc . The colors associated with all the edges that are not incident with vi or
vc , or the edges that join vertices in C = C ∗ ∪ {vc } remain the same. An edge {vi , w}, w ∈ /C
receives color σ1 if it has color σ1 or σ3 with respect to ∆∗ , otherwise it is assigned σ2 . On the
other hand, an edge {vc , w}, w ∈ / C receives color σ3 if {vi , w} has color σ2 or σ3 with respect
∗
to ∆ , otherwise it is assigned color σ4 . It is easy to see that ∆ has no monochromatic I` .
The crucial fact about this construction is that the color of {vi , w} that was originally
assigned by ∆∗ is uniquely determined by the colors of {vi , w} and {vc , w} in ∆ = φ(∆∗ ).
The injectivity of φ is an easy consequence of this fact, and hence
X X
|∆s∗ (G∗ )| ≤ |∆s (G)| .
s∗ ∈S1 s∈S2
Now, note that we may interchange the roles of the colors in the above mapping, that is, the
colors σ1 , σ2 could be assigned to vc , while σ3 , σ4 would be assigned to vi . Since all the colors
are substantial for vi with respect to the colorings in ∆s∗ (G∗ ), the colorings created with this
new mapping are all distinct from the colorings created through the original mapping. By
the argument above we get
X X
2 |∆s∗ (G∗ )| ≤ |∆s (G)| ,
s∗ ∈S1 s∈S2

which establishes (17) and thus concludes the proof. 

Acknowledgements: The authors are indebted to an anonymous referee for valuable comments
and suggestions. The authors are also grateful to NUMEC/USP, Núcleo de Modelagem Es-
tocástica e Complexidade of the University of São Paulo, for its hospitality.
 EDGE COLORINGS AVOIDING MONOCHROMATIC MATCHINGS OF A GIVEN SIZE 13

References
1. N. Alon, J. Balogh, P. Keevash, and B. Sudakov, The number of edge colorings with no monochromatic
cliques, J. London Math. Soc. (2) 70, 2004, 273–288.
2. J. Balogh, A remark on the number of edge colorings of graphs, European J. Combin. 77, 2006, 565–573.
3. P. Erdős, Some new applications of probability methods to combinatorial analysis and graph theory, Proc. of
the Fifth Southeastern Conference on Combinatorics, Graph Theory and Computing (Florida Atlantic
Univ., Boca Raton, Fla., 1974), 39–51. Congressus Numerantium, No. X, Utilitas Math., 1974, 39–51.
4. P. Erdős and T. Gallai, On maximal paths and circuits, Acta Math. Acad. Sci. Hungar. 10, 1959, 337–356.
5. P. Erdős and A. H. Stone, On the structure of linear graphs, Bull. Amer. Math. Soc. 52, 1946, 1087–1091.
6. C. Hoppen, Y. Kohayakawa, and H. Lefmann, Hypergraphs with many Kneser colorings, 39 pp., short
version: Kneser Colorings in Uniform Hypergraphs, Electronic Notes in Discrete Mathematics 34, 2009,
219–223, Proceedings of EUROCOMB 2009.
7. C. Hoppen, Y. Kohayakawa, and H. Lefmann, Edge colorings of graphs avoiding some fixed monochromatic
subgraph with linear Turán number, to appear in Electronic Notes in Discrete Mathematics, Proceedings
of EUROCOMB 2011.
8. H. Lefmann and Y. Person, Exact results on the number of restricted edge colorings for some families of
linear hypergraphs, submitted.
9. H. Lefmann, Y. Person, V. Rödl, and M. Schacht, On colorings of hypergraphs without monochromatic
Fano planes, Combinatorics, Probability & Computing 18, 2009, 803–818.
10. H. Lefmann, Y. Person, and M. Schacht, A structural result for hypergraphs with many restricted edge
colorings, Journal of Combinatorics 1, 2010, 441–475.
11. O. Pikhurko, and Z. B. Yilma, The maximum number of K3 -free and K4 -free edge 4-colorings, to appear
in J. London Math. Soc., 24 pp.
12. E. Szemerédi, Regular partitions of graphs, Problèmes combinatoires et théorie des graphes (Colloq. In-
ternat. CNRS, Univ. Orsay, Orsay, 1976), Colloq. Internat. CNRS, vol. 260, CNRS, Paris, 1978, 399–401.
13. R. Yuster, The number of edge colorings with no monochromatic triangle, J. Graph Theory 21, 1996,
441–452.

Instituto de Matemática, Universidade Federal do Rio Grande do Sul, Avenida Bento Gonçalves,
9500, 91509-900 Porto Alegre, Brazil
E-mail address: [email protected]

Instituto de Matemática e Estatı́stica, Universidade de São Paulo, Rua do Matão, 1010,

05508-090, São Paulo, Brazil
E-mail address: [email protected]

Fakultät für Informatik, Technische Universität Chemnitz, Straße der Nationen 62, D-09107
Chemnitz, Germany
E-mail address: [email protected]